hIRT 0.5.0
- The computational core is now written in C++ (Rcpp/RcppArmadillo):
the E-step, the item-level M-step (a Newton solver for the weighted
cumulative logit model, replacing
rms::lrm.fit() and
glm.fit() inside the EM loop), the level-II variance
regression, and the scores used for standard errors. Estimates, standard
errors, and log-likelihoods agree with version 0.4.0 to numerical
precision, and fits are roughly 10 to 100 times faster.
- The E-step is computed on the log scale, so it no longer underflows
when a respondent answers many items.
hgrmDIF() pools respondents who share the same DIF
covariate pattern in the M-step, which makes it much faster when
x0 is categorical.
hIRT 0.4.0
- Removed dependency on the archived package pryr; function
composition is now done in base R.
- Added
hgrmDIF() for differential item functioning in
the graded response model.
hIRT 0.3.1